| Guide | GDE-087 |
| Topic | AI text detection |
| Domain | AI-assisted investigation |
| Last tested | 29 October and 1 November 2026 |
Short answer: To check whether text was written by AI, preserve and date it, run it through two AI detectors, then verify every citation it contains. An AI detector score is a probability, not proof, so never report it as a verdict.
What is an AI detector?
An AI detector is a tool that estimates whether a piece of text was written by a large language model such as ChatGPT, Gemini or Claude, by scoring statistical patterns in word choice and sentence structure. For investigators, its score is one signal to weigh alongside provenance, dating and source checks, never a verdict on its own.
AI detectors are classifiers trained on large samples of human and machine writing. They return a probability or a percentage, often with highlighted sentences, rather than proof. OpenAI withdrew its own text classifier in July 2023, citing its low rate of accuracy, after it correctly identified only 26% of AI-written text in its own evaluation. A Stanford-led study published in Patterns found that seven widely used detectors labelled more than half of 91 TOEFL essays by non-native English writers as AI-generated.
Used well, detection sits inside a wider workflow: establish when and where the text first appeared, test it on more than one detector, then look for positive evidence such as fabricated citations or leftover chatbot phrasing.
| • | Checking whether a statement or open letter attributed to a named person was machine-written |
| • | Assessing a tip, leak or witness account that reads as generic or unusually polished |
| • | Reviewing a batch of reviews, comments or posts suspected of being mass-produced |
| • | Testing whether a high-volume website publishes generated articles |
| • | Weighing a document's claimed date against when AI writing tools became widely available |
How do you check if text was written by AI?
Preserve the text, date it, then test it on more than one detector and look for the evidence a detector cannot see.
Five tools cover the workflow, each with a free route in.
Google Search: Free, no login required. Exact-phrase searches in quotation marks find earlier copies of the same wording, which can reveal a human original or a shared template.
Wayback Machine: Free, no login required to search. The Internet Archive's capture history shows when a crawler first saw a page and its text, which helps you date the writing.
GPTZero: Free and paid tiers. Returns a percentage estimate with highlighted passages. Scans of up to 10,000 characters work without an account; longer scans need a free account.
Copyleaks AI Detector: Free and paid tiers. Free scans of up to 25,000 characters are offered in more than 30 languages. A free account unlocks more monthly credits.
Pangram: Free and paid tiers. Classifies text as human, AI or AI-assisted, with segment-level analysis and support for more than 20 languages. Free checks need a free account.
What we tested: On 1 November 2026 we ran Google Search, the Wayback Machine and GPTZero on two test passages: 556 characters of the public-domain Gettysburg Address and a 491-character passage we wrote ourselves in a generic machine-like style. No private or personal text was used, and the published figures were checked against their sources on 29 October 2026.
Google Search, the Wayback Machine and GPTZero returned results on every step we ran. Copyleaks AI Detector was Not tested because its Scan button says scanning means agreeing to its Terms of use, and Pangram was Not tested because it needs an account. Our browser was already signed in to Google and GPTZero, so we could not confirm that either works without an account, and the Copyleaks no-account and sentence-level claims could not be confirmed from its page.
On 1 November 2026 an exact-phrase search for a line of the public-domain Gettysburg Address returned earlier copies on Wikipedia, the Library of Congress and other sites (1 of 1). Our browser was already signed in to Google, so we could not confirm that no sign-in is needed. The Advanced Search page on 29 October 2026 loaded with an exact-phrase field and no sign-in prompt. | ||
On 1 November 2026 the capture calendar for a public page carrying the Gettysburg Address loaded with no sign-in and reported 2,504 captures between 6 February 2013 and 23 October 2026 (1 of 1). On 29 October 2026 its availability endpoint returned a capture for openai.com and none for example.com (1 of 2 lookups); we did not rerun it. | ||
On 1 November 2026 two passages were scanned from the home page (2 of 2 scans ran). The 556-character public-domain Gettysburg Address returned 100% human and 0% AI. A 491-character passage we wrote in a generic machine-like style returned 100% AI, with a notice that text under 100 words may be less accurate. Our browser was already signed in to a GPTZero account, so we could not confirm that scans work without one. The live page on 29 October 2026 states a 10,000-character limit for scans and asks for a free account above it. | ||
On 1 November 2026 the page loaded and accepted pasted text with no sign-in, but its Scan button states that scanning means agreeing to Copyleaks' Terms of use, so we did not scan and cleared the text. The live page on 29 October 2026 states free scans up to 25,000 characters and more than 30 languages, and says a free account unlocks more monthly credits. It does not state whether sign-in is needed or whether results are sentence-level, so those two claims are unconfirmed. | ||
Its FAQ states 20 free text checks a day after signing up and a human, AI or AI-assisted prediction. It needs an account, so no check was run. | ||
OpenAI's page states 26% of AI-written text correctly flagged and the classifier withdrawn on 20 July 2023. The Patterns paper on arXiv states seven detectors, 91 TOEFL essays and a 61.22% average false positive rate. |
This is a snapshot of one search, one calendar lookup and two short passages on one detector, with the other tools read from public pages on 29 October 2026, and not a measure of how accurate any detector is.
The Second-Strand Rule: Never call text machine-written or human-written on a detector score alone. Support the call with one finding from outside the detector, such as a dated earlier copy or a fabricated citation.